Look Before You Edit: Attention-Guided Camera Placement and Multi-View Alignment for 3D Gaussian Splatting Editing

2026-07-22

AI Summary

Existing text-driven 3D Gaussian splatting editing methods are constrained by fixed training viewpoints, struggling to simultaneously achieve local precision and multi-view consistency. This work proposes an attention-guided editing framework that first leverages self- and cross-attention maps from diffusion models to dynamically select optimal editing viewpoints. It then constructs a 3D attention field, enabling token-level feature sharing and multi-view attention alignment to jointly optimize both edit locality and consistency across appearance and spatial dimensions. Requiring only five editing views, the proposed method outperforms existing approaches in instruction fidelity, multi-view consistency, and edit locality, yielding significantly higher user preference and reducing editing latency by up to 7×.

Abstract

Text-driven 3D scene editing with 3D Gaussian Splatting (3DGS) typically applies a 2D diffusion editor to views rendered from fixed training cameras, limiting both the spatial coverage of edits and the user's freedom to target specific objects in complex scenes. We present LB-Edit, a framework that addresses two coupled problems: where to place editing cameras for localized edits, and how to make per-view edits agree with one another so that the 3D scene remains consistent after fine-tuning. First, Attention-Guided Editing Camera Placement (ACP) probes the diffusion model's self- and cross-attention at multiple candidate camera distances to find where attention is well-contained in the region of interest, then places a compact, geometrically diverse editing camera set at that attention-optimal distance. Second, Multi-view Attention Alignment (MAA) steers the editor toward the same edit across views along two axes: it aligns appearance by sharing self-attention features via token-level correspondence, and aligns spatial location by lifting cross-attention maps onto the 3D Gaussians as a shared 3D attention field, suppressing both appearance and spatial drift. Experiments on multi-object and single-object scenes show that our method achieves the highest user preference in instruction fidelity, multi-view consistency, and editing locality, using as few as 5 editing views and reducing latency by up to 7x over existing methods.

Problem

• 3D Gaussian Splatting
• text-driven editing
• camera placement
• multi-view consistency
• 3D scene editing

Innovation

• Attention-Guided Camera Placement
• Multi-view Attention Alignment
• 3D Gaussian Splatting Editing
• Diffusion-based 3D Editing
• 3D Consistency